Perks, O. F. J., Hammond, Simon D., Pennycook, Simon J. and Jarvis, Stephen A. (2011) WMTrace - a lightweight memory allocation tracker and analysis framework. In: UK Performance Engineering Workshop (UKPEW'11), Bradford, United Kingdom, 7-8 July, 2011

Abstract

The diverging gap between processor and memory performance has been a well discussed aspect of computer architecture literature for some years. The use of multi-core processor designs has, however, brought new problems to the design of memory architectures - increased core density without matched improvement in memory capacity is reduc- ing the available memory per parallel process. Multiple cores accessing memory simultaneously degrades performance as a result of resource con- tention for memory channels and physical DIMMs. These issues combine to ensure that memory remains an on-going challenge in the design of parallel algorithms which scale. In this paper we present WMTrace, a lightweight tool to trace and analyse memory allocation events in parallel applications. This tool is able to dynamically link to pre-existing application binaries requiring no source code modification or recompilation. A post-execution analysis stage enables in-depth analysis of traces to be performed allowing memory allocations to be analysed by time, size or function. The second half of this paper features a case study in which we apply WMTrace to five parallel scientific applications and benchmarks, demonstrating its effectiveness at recording high-water mark memory consumption as well as memory use per-function over time. An in-depth analysis is provided for an unstructured mesh benchmark which reveals significant memory allocation imbalance across its participating processes.

Request changes to a record

Actions (login required)

View Item